[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7FDS4FXMAECFNPBAAZ6ZMC6":3,"skill-same:01M7FDS4FXMAECFNPBAAZ6ZMC6":64},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7FDS4FXMAECFNPBAAZ6ZMC6","lean-formalize",null,"Develop and verify a mathematical proof in Lean, continue an incomplete Lean project, or audit whether it proves the original statement. Connect actual inputs to intermediate lemmas, assemble the target theorem, check its transitive axioms, and provide a reproducible handoff. Use when Lean is requested or a specific proof obligation benefits from formal verification; use proof-writer for ordinary mathematical drafting.","CZX-LBJ\u002Fdeepseek-science","CZX-LBJ","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F204594689?v=4","01M7FC59WQ9PFXT04MKGJDCFFT","czx","01M7FC59XB2GW0BKWN9JTDBQS7",1,"MIT","https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fskills-codex\u002Flean-formalize\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: lean-formalize\ndescription: \"Develop and verify a mathematical proof in Lean, continue an incomplete Lean project, or audit whether it proves the original statement. Connect actual inputs to intermediate lemmas, assemble the target theorem, check its transitive axioms, and provide a reproducible handoff. Use when Lean is requested or a specific proof obligation benefits from formal verification; use proof-writer for ordinary mathematical drafting.\"\n---\n\n# Lean Formalize\n\nTurn the user's mathematical statement into a checked Lean theorem with its\nmeaning preserved. A compiled conditional lemma is progress; completion concerns\nthe original statement and the trust basis actually used.\n\n## When to use Lean\n\nUse this skill when the user requests Lean, when continuing an existing Lean\nproof, or when formal verification addresses a concrete uncertainty in a central\nclaim—for example, a long dependency chain, a delicate reduction, or coverage of\na finite classification. State the obligation it will help resolve and proceed\nwithin the authorized task. Difficulty alone is not a reason to formalize.\nOrdinary derivations and short proofs can stay in `formula-derivation` or\n`proof-writer`; do not make Lean a prerequisite for every mathematical result.\nRespect the user's chosen proof method and the scale of the requested work.\n\n## Core workflow\n\n```text\nOriginal statement and Lean definitions\n  → A: cross-family adversarial statement alignment\n  → Proof obligations, representations, and lemma interfaces\n  → Lean implementation ↔ B: adversarial review of key arguments and connections\n  → Actual inputs connected; original theorem assembled\n  → Executed type, definition, and transitive-axiom audit\n  → C: cross-family adversarial review of the final exported result\n  → Reproducible delivery and research-state update\n```\n\nFor substantial new proof projects, A\u002FB\u002FC are part of the workflow. For a\ncontinuation, reuse completed checks on unchanged claims and revisit affected\nones. Small routine formalizations need checks proportional to the actual claim;\nan explicit user request for cross-family review still applies to them.\n\nUse the authorized reviewer families available in the current host. If the user\nspecifies both Grok and Gemini, obtain and record both; a same-family agent or\nanother provider does not silently satisfy either request. Unavailability leaves\nthat checkpoint pending while independent proof work continues. A checked theorem\nand a fully completed requested review workflow are separate deliverables.\n\n## Scope and entry\n\nFollow the requested scope: implement, continue, audit, or package. For an audit,\ninspect and report rather than silently repairing or weakening the theorem.\nFor implementation, develop the mathematical argument as well as its Lean\nproof: search relevant literature or library results, explore alternative routes,\nand discharge the missing lemmas. Continue through the remaining interfaces and\ntop-level assembly; do not stop at a target declaration or the first successful\ncompilation.\n\nRead the authoritative statement, existing Lean entry point, toolchain and\nlockfile, and latest progress record. Reuse the project and its document names.\nDo not reinstall tools, change dependencies, or start a new proof framework when\nthe existing environment is suitable. Resolve APIs against the pinned library.\n\nUse the existing proof route when it works. If the mathematical argument itself\nis missing, isolate that obligation and use `proof-writer` or ordinary proof work.\nA delegated `proof-writer` task develops that argument and returns its proof or\nremaining gap to this run; it must not invoke `lean-formalize` again. Syntax\nautomation cannot discharge an unproved premise. Do not promise that an arbitrary\nopen problem can be formalized or solved.\n\n### Start or resume the right work\n\n| Current input | First useful action |\n|---|---|\n|Only a mathematical statement|Fix definitions and quantifiers, then develop a proof route and its first difficult obligation|\n|A prose proof|Identify nontrivial dependencies and choose Lean representations; expose any missing argument before coding it|\n|A partial Lean project|Inspect the target and its actual callers, read the last useful build\u002Ferror record, and continue at the highest unclosed connection|\n|An audit request|Read definitions and exported types, run applicable checks, and report; do not silently repair the target|\n|A completed proof to hand off|Check the current entry point and evidence, then prepare portable sources and commands without restarting the mathematical search|\n\nName the intended main module, exported declaration, and current next obligation\nearly. If no complete mathematical route is known, say which statement is being\nattempted; do not mark it provable merely because the implementation has started.\nTool\u002FAPI problems and missing mathematical arguments require different next steps.\n\n## 1. Fix meaning before implementation\n\nRecord a short mathematical specification, or reference the existing one:\n\n- Objects, domains, quantifiers, original hypotheses, and conclusion.\n- Equivalence of convenient representations to the original objects.\n- User constraints on computation, external certificates, and logical foundations.\n- The Lean declarations intended to express and prove the result.\n\nSeparate original hypotheses from properties introduced by a reduction. Identify\nwhere finiteness, nonemptiness, decidability, normalization, and index conventions\nchange the statement or require a bridge. Check plausible vacuity and quantifier\nfailures in the actual theorem, rather than inventing unrelated edge cases.\n\nKeep the user's original statement as the comparison baseline until the user\nchanges the goal. A working specification rewritten to match the implementation\ndoes not change that baseline. Record authorized scope changes explicitly; do not\nrequest confirmation again for a change already authorized in the session.\nIf a repair yields only a stronger assumption or weaker conclusion, identify the\nproved variant and the original obligation still open. A stronger proved result\ncan establish the original claim when its implication is supplied.\n\nRun checkpoint A on the actual definition bodies and proposed target. Ask for\nindependent back-translation before comparison with the original mathematics.\nThe input packet and prompt are in\n[references\u002Fadversarial-review.md](references\u002Fadversarial-review.md). A reviewer\nwho saw only the intended prose has not checked the encoding.\n\n## 2. Design the connections, then build useful pieces\n\nMap the path from an arbitrary original input to the conclusion. For every\nsubstantial interface, record:\n\n| Declaration \u002F obligation | What it assumes | Where actual inputs come from | Evidence \u002F remaining gap |\n|---|---|---|---|\n|A conditional result|Its extra hypotheses|A named construction or theorem from the original input|Actual state|\n\nPrioritize the highest unclosed connection. In particular, distinguish:\n\n1. A formula or certificate computes the desired number.\n2. The actual mathematical object realizes that formula or certificate.\n3. The computed fact implies the original conclusion.\n\nAll three may require separate proofs. A structure that stores its desired\nproperties as fields, a supplied probability bound, or an assumption equivalent\nto the conclusion does not remove the obligation to construct that input.\n\nUse conditional lemmas as development interfaces, explicitly marked as such.\nKeep incomplete experiments outside the certified target's dependency chain;\nany temporary `sorry` remains visible as unfinished work and cannot survive the\nfinal target audit. Prefer enough intermediate compilation to localize errors,\nwithout interpreting file counts or proved arithmetic statements as completion.\n\nSplit parallel work along stable lemma signatures and module ownership. Give\neach worker its assumptions, conclusion, dependencies, and concrete compile\ntarget. Integrate its result against the actual caller before closing the ledger.\nDo not let parallel workers silently redefine shared objects to suit their proofs.\n\n### Implementation loop\n\nRead [references\u002Flean-working-loop.md](references\u002Flean-working-loop.md) when\nimplementing or repairing an obligation. It develops the search → actual-caller\nexperiment → diagnostic → repair cycle, including representation choices and\nperformance problems, with a compiled library-application example.\n\n1. Select a missing connection or mathematical lemma that changes what the main\n   theorem can prove. State its exact Lean interface and its caller's obligations.\n2. Look for the required results in the pinned library and project. Test unfamiliar\n   declarations locally with `#check`; use existing equivalent representations\n   when they simplify a real bottleneck.\n3. Implement the lemma and an actual use site. Compile the affected module; after\n   integration, compile the downstream target whose status depends on it.\n4. Diagnose the first relevant error. An elaboration\u002FAPI mismatch calls for a\n   local implementation fix. An unavailable assumption calls for its derivation,\n   a different argument, or an explicit remaining mathematical obligation.\n5. When finite reduction becomes expensive, consider a general counting lemma,\n   recurrence or smaller checked certificate. Do not silently enlarge the trust\n   basis merely to make a tactic finish.\n6. At a new load-bearing argument or interface, run checkpoint B. Implement valid\n   fixes, compile them, and request follow-up on the changed obligation.\n7. Update the existing ledger with the declaration, actual caller, verification\n   evidence and remaining premise; then advance to the next connection.\n\nFor example, a theorem of type `Certificate x → Desired x` is useful only after\nthe project constructs `Certificate x` for every original input it needs. Closing\nthat construction and connecting the caller is a distinct result from proving\nthe conditional theorem. Kernel checking does not discharge a parameter merely\nbecause its type has a reassuring name.\n\nDo not use a timer, repeated unchanged builds, or repeated model calls as a proxy\nfor progress. After a concrete failure, try another justified representation or\nproof path; preserve the exact blocker if the requested work cannot yet finish.\nDistinguish an implementation failure from a failed intermediate claim and an\nobstruction to the whole method. Retire an unsuccessful route with its reason;\nrejecting that route does not refute the original theorem. Parallel exploration\nis most useful when the proposed approaches can fail for different reasons.\n\n## 3. Use computation with a proved interpretation\n\nDistinguish finite examples, exhaustive computation over a proved domain,\nchecked certificates, and symbolic\u002Fgeneral proof. A bounded search is evidence\nabout its searched inputs. Exhaustive finite verification can be a proof when\ncoverage, encoding correspondence, and the verification procedure are established\nand the user's constraints permit it. For an end-to-end Lean result, the coverage\nand interpretation must themselves lie in the proved dependency chain under the\ndeclared foundations. A comment claiming that a finite list is exhaustive does\nnot turn checked list entries into a universal Lean theorem.\n\nFor numeric arguments, use exact arithmetic where the claim needs exactness.\nProve the connection between actual objects\u002Fevents and the finite data before\nusing the numeric conclusion. Document conventions that matter: ordered versus\nunordered pairs, multiplicities, indices, zero cases, and rounding.\n\nAgree on the intended trust basis from the specification and project conventions.\nDo not silently introduce axioms or compiler-backed computation to make an\notherwise incomplete kernel-level proof appear complete. Conversely, do not turn\none project's prohibition on enumeration or `native_decide` into a universal\nban. Explain any extra trust assumption and whether it satisfies this task.\n\n## 4. Review the mathematical obligations that remain uncertain\n\nRun A, B and C as described in the core workflow. Read\n[references\u002Fadversarial-review.md](references\u002Fadversarial-review.md) for stage\ntriggers, source packets, adversarial prompts, objection handling, resumption and\nstopping rules. These are mathematical checks, not votes on whether to trust the\nexecutor's summary.\n\nUse `proof-checker` for the existing ARIS proof-audit\u002Fsubmission role. This skill\ndoes not replace its artifact or reviewer contract. When invoked by a parent\nproof audit, return the checked scope and evidence to that run; do not recursively\ninvoke `proof-checker`. A standalone Lean task does not automatically require a\nseparate paper-audit workflow. Where installed, use\n`research-review` \u002F `auto-review-loop` for targeted discussion and repair, with\nthe current host's reviewer routing.\n\nFor direct model consultations, use available authorized tools and their actual\ncontracts. A Grok or Antigravity MCP consultation is not automatically an ARIS\nreviewer overlay. Do not hard-code vendor availability, quotas, retry counts,\nproxy settings, or a particular model into the portable workflow.\n\nAsk concrete questions: translate this declaration; discharge these caller\nhypotheses; find a counterexample to this new lemma; check this reduction's\ncoverage. Preserve raw responses, actual reading\u002Fexecution scope, and the\nexecutor's disposition of each issue. Model agreement is not a proof. Same-family\nopinions remain provisional; cross-family opinions are additional review\nevidence, not mathematical truth certificates.\n\nAn unread artifact, an unanswered question, a failed call or an empty response\nprovides no substantive review verdict. A plan review is not a final-code review;\nreading existing logs is not executing a build. Preserve which claim\u002Fversion was\nactually examined. Retry only within current authorization and provider rules;\ndo not rewrite failures as passes after a later success.\n\nClose objections with a specific derivation, verified witness, correction or\njustified rejection. Preserve unresolved disputes. Apply an invoked ARIS loop's\nown round\u002Fcontinuation contract; do not create an outer polling loop around it.\nFor direct consultations, re-review changed obligations and stop when settled or\nwhen a real availability\u002Fbudget limit is reached. Unresolved mathematics remains\nopen; an unavailable requested review remains pending. Neither a reviewer score\nnor exhaustion of rounds is a proof-completion rule.\n\n### Host execution and review\n\nRead the local host's [reviewer routing](..\u002Fshared-references\u002Freviewer-routing.md)\nfor its current model\u002Feffort policy and tool contracts. Use the A\u002FB\u002FC mathematical\npackets in this skill; they do not require a paper, publication score, or a new\nreview service. For delegated proof work, follow the applicable\n[fan-out convention](..\u002Fshared-references\u002Ffan-out-pattern.md), assign stable lemma\ninterfaces, and integrate each result at its real caller.\n\nUse Codex's available native shell, file, and collaboration tools for Lean work.\nA second Codex agent is useful for decomposition, implementation, or provisional\nreview; record `review_independence: same-family` and\n`acceptance_status: provisional`. It does not fulfill a requested cross-family\ncheckpoint.\n\nFor A\u002FB\u002FC, use the user's selected authorized cross-family reviewer, or an\nalready configured cross-family route: Claude review MCP, Grok MCP, or Gemini\nthrough its review bridge or Antigravity. An installed `research-review` overlay\nmay carry the packet if its review contract fits; the base Codex version remains\nsame-family. Existing overlays cover only their listed skills and do not\nimplicitly override `lean-formalize`. Direct checkpoint calls need no new overlay;\nread [review transport notes](references\u002Fadversarial-review.md#review-transport-and-source-access)\nfor the actual file-access and continuation rules. Do not invent new global\nreviewer flag values. If no cross-family route is available, continue useful\nlocal Lean work and record the affected checkpoint as pending.\n\n## 5. Verify the actual final target\n\nUse [references\u002Fverification-and-handoff.md](references\u002Fverification-and-handoff.md)\nwhen preparing the final build, trust audit, or portable package.\n\nFor completed delivery, establish the applicable evidence below. Keep a proved\ntheorem distinct from any still-pending reproduction or review requirement:\n\n- **The proposition has a proof:** inspect the exported declaration's actual\n  type, not its keyword. `theorem`, `lemma`, or `def` may supply a proof term of\n  the target proposition. Merely defining that proposition, proving `P → P`, or\n  proving an unrelated proposition does not discharge the original target.\n- **Statement alignment:** its expanded type and definitions express the original\n  claim; no internal profile, certificate-validity, coverage, or success premise\n  is left for the user to supply unless the original claim included it.\n- **Actual assembly:** reductions, cases, witnesses, and return to the original\n  conclusion are connected, with all caller hypotheses discharged.\n- **Executed verification:** explicitly build the module that proves the target\n  and print its type and transitive axioms. For a multi-module project, also\n  independently import the actual entry point to check its exported result.\n  Direct compilation plus type\u002Faxiom output suffices for a small self-contained\n  file; do not build a new project just to create another import. Save commands,\n  exit status, and output. A successful build of a different root module does\n  not count.\n- **Trust accounting:** no `sorryAx` or unapproved extra axiom lies on the target's\n  dependency chain. Explain the actual dependencies against the intended basis;\n  the required list need not be exactly any fixed set of axiom names.\n- **Reproducibility:** retain the relevant source and pinned environment, and\n  provide commands that another reader can run.\n- **Requested review:** C concerns the final exported type, definition bodies,\n  axiom output and original claim, not an earlier plan. Record each required\n  reviewer's substantive findings or unavailability separately.\n\nThese checks establish the formal result under its stated foundations. A prose\nproof or HTML edition requires its own correspondence check; Lean does not certify\nevery sentence in a separate exposition. An explicitly required import audit,\nexternal reproduction, or review may remain pending even after the formal theorem\nis checked. Report that missing evidence as such, not as a newly discovered\nmathematical gap.\n\nAn illustrative audit can distinguish `available_interface : Target → Target`\nfrom `original_proof : Target`, even if both compile without extra axioms. The\nfirst leaves the original claim open; the second supplies its proof when `Target`\nmatches the user's statement. A failed reviewer call changes neither type.\n\n## 6. Persist and hand off the evidence actually obtained\n\nReuse the existing specification, proof ledger, and continuation file. A short\nprogress report plus Lean sources, lockfiles, an audit entry point, and logs is\nusually enough. Add new files only for a concrete reader or execution need.\n\nReport distinct facts rather than a single ambiguous PASS:\n\n- Original theorem: checked, partial, changed claim only, or disproved by a\n  verified counterexample; name the declaration and any remaining obligation.\n- Statement alignment and the actual logical trust basis.\n- Reproduction: existing\u002Fincremental build, clean project rebuild, or another\n  person's independent reproduction—whichever actually happened.\n- Reviews: who read what, who ran code, and any requested review still pending.\n\nFor unfinished work, save the smallest remaining mathematical\u002FLean obligation,\nits known dependencies, the last useful error, and the next action. For completed\nwork, record completion so the next session does not revive obsolete gaps.\n\nUse a compact continuation note rather than a second state system:\n\n```text\nOriginal target and source:\nCurrent exported declaration \u002F main module:\nClosed connections and actual build evidence:\nNext unclosed obligation or completion:\nA\u002FB\u002FC reviews: source scope, reviewer, raw record, remaining issue:\nNext action:\n```\n\nResolve current state from actual artifacts, not an old narrative saying either\n“finished” or “not finished.” On resume, check what changed before repeating work.\nDo not merge historical and current reviewer scopes into a larger claimed audit.\n\nIf a research wiki is active, update the affected claim through the existing\n`research-wiki` \u002F `proof-checker` workflow with precise evidence and scope; do not\noverwrite its generated graph or invent a second acceptance format. Literature\nsearch should resolve a mathematical gap, API question, or relevant prior result,\nnot become an unrelated requirement to exhaust the literature.\n\nFor handoff, provide a reader-oriented proof, source, pinned dependencies,\nindependent audit command, and known trust boundary. Use `render-html` when a\nsingle-file reading edition would help. Mathematical source changes require\nappropriate rebuild and statement\u002Faxiom checks. A theorem or hypothesis changed\nin Markdown or HTML is also a claim change: redo alignment and affected requested\nreview. Layout or wording that preserves the claim needs presentation checks.\nKeep change histories in the research record and present the finished proof\ndirectly. Do not restart completed proof work solely to satisfy a workflow ritual.\n","# Lean Formalize\n\nTurn the user's mathematical statement into a checked Lean theorem with its\nmeaning preserved. A compiled conditional lemma is progress; completion concerns\nthe original statement and the trust basis actually used.\n\n## When to use Lean\n\nUse this skill when the user requests Lean, when continuing an existing Lean\nproof, or when formal verification addresses a concrete uncertainty in a central\nclaim—for example, a long dependency chain, a delicate reduction, or coverage of\na finite classification. State the obligation it will help resolve and proceed\nwithin the authorized task. Difficulty alone is not a reason to formalize.\nOrdinary derivations and short proofs can stay in `formula-derivation` or\n`proof-writer`; do not make Lean a prerequisite for every mathematical result.\nRespect the user's chosen proof method and the scale of the requested work.\n\n## Core workflow\n\n```text\nOriginal statement and Lean definitions\n  → A: cross-family adversarial statement alignment\n  → Proof obligations, representations, and lemma interfaces\n  → Lean implementation ↔ B: adversarial review of key arguments and connections\n  → Actual inputs connected; original theorem assembled\n  → Executed type, definition, and transitive-axiom audit\n  → C: cross-family adversarial review of the final exported result\n  → Reproducible delivery and research-state update\n```\n\nFor substantial new proof projects, A\u002FB\u002FC are part of the workflow. For a\ncontinuation, reuse completed checks on unchanged claims and revisit affected\nones. Small routine formalizations need checks proportional to the actual claim;\nan explicit user request for cross-family review still applies to them.\n\nUse the authorized reviewer families available in the current host. If the user\nspecifies both Grok and Gemini, obtain and record both; a same-family agent or\nanother provider does not silently satisfy either request. Unavailability leaves\nthat checkpoint pending while independent proof work continues. A checked theorem\nand a fully completed requested review workflow are separate deliverables.\n\n## Scope and entry\n\nFollow the requested scope: implement, continue, audit, or package. For an audit,\ninspect and report rather than silently repairing or weakening the theorem.\nFor implementation, develop the mathematical argument as well as its Lean\nproof: search relevant literature or library results, explore alternative routes,\nand discharge the missing lemmas. Continue through the remaining interfaces and\ntop-level assembly; do not stop at a target declaration or the first successful\ncompilation.\n\nRead the authoritative statement, existing Lean entry point, toolchain and\nlockfile, and latest progress record. Reuse the project and its document names.\nDo not reinstall tools, change dependencies, or start a new proof framework when\nthe existing environment is suitable. Resolve APIs against the pinned library.\n\nUse the existing proof route when it works. If the mathematical argument itself\nis missing, isolate that obligation and use `proof-writer` or ordinary proof work.\nA delegated `proof-writer` task develops that argument and returns its proof or\nremaining gap to this run; it must not invoke `lean-formalize` again. Syntax\nautomation cannot discharge an unproved premise. Do not promise that an arbitrary\nopen problem can be formalized or solved.\n\n### Start or resume the right work\n\n| Current input | First useful action |\n|---|---|\n|Only a mathematical statement|Fix definitions and quantifiers, then develop a proof route and its first difficult obligation|\n|A prose proof|Identify nontrivial dependencies and choose Lean representations; expose any missing argument before coding it|\n|A partial Lean project|Inspect the target and its actual callers, read the last useful build\u002Ferror record, and continue at the highest unclosed connection|\n|An audit request|Read definitions and exported types, run applicable checks, and report; do not silently repair the target|\n|A completed proof to hand off|Check the current entry point and evidence, then prepare portable sources and commands without restarting the mathematical search|\n\nName the intended main module, exported declaration, and current next obligation\nearly. If no complete mathematical route is known, say which statement is being\nattempted; do not mark it provable merely because the implementation has started.\nTool\u002FAPI problems and missing mathematical arguments require different next steps.\n\n## 1. Fix meaning before implementation\n\nRecord a short mathematical specification, or reference the existing one:\n\n- Objects, domains, quantifiers, original hypotheses, and conclusion.\n- Equivalence of convenient representations to the original objects.\n- User constraints on computation, external certificates, and logical foundations.\n- The Lean declarations intended to express and prove the result.\n\nSeparate original hypotheses from properties introduced by a reduction. Identify\nwhere finiteness, nonemptiness, decidability, normalization, and index conventions\nchange the statement or require a bridge. Check plausible vacuity and quantifier\nfailures in the actual theorem, rather than inventing unrelated edge cases.\n\nKeep the user's original statement as the comparison baseline until the user\nchanges the goal. A working specification rewritten to match the implementation\ndoes not change that baseline. Record authorized scope changes explicitly; do not\nrequest confirmation again for a change already authorized in the session.\nIf a repair yields only a stronger assumption or weaker conclusion, identify the\nproved variant and the original obligation still open. A stronger proved result\ncan establish the original claim when its implication is supplied.\n\nRun checkpoint A on the actual definition bodies and proposed target. Ask for\nindependent back-translation before comparison with the original mathematics.\nThe input packet and prompt are in\n[references\u002Fadversarial-review.md](references\u002Fadversarial-review.md). A reviewer\nwho saw only the intended prose has not checked the encoding.\n\n## 2. Design the connections, then build useful pieces\n\nMap the path from an arbitrary original input to the conclusion. For every\nsubstantial interface, record:\n\n| Declaration \u002F obligation | What it assumes | Where actual inputs come from | Evidence \u002F remaining gap |\n|---|---|---|---|\n|A conditional result|Its extra hypotheses|A named construction or theorem from the original input|Actual state|\n\nPrioritize the highest unclosed connection. In particular, distinguish:\n\n1. A formula or certificate computes the desired number.\n2. The actual mathematical object realizes that formula or certificate.\n3. The computed fact implies the original conclusion.\n\nAll three may require separate proofs. A structure that stores its desired\nproperties as fields, a supplied probability bound, or an assumption equivalent\nto the conclusion does not remove the obligation to construct that input.\n\nUse conditional lemmas as development interfaces, explicitly marked as such.\nKeep incomplete experiments outside the certified target's dependency chain;\nany temporary `sorry` remains visible as unfinished work and cannot survive the\nfinal target audit. Prefer enough intermediate compilation to localize errors,\nwithout interpreting file counts or proved arithmetic statements as completion.\n\nSplit parallel work along stable lemma signatures and module ownership. Give\neach worker its assumptions, conclusion, dependencies, and concrete compile\ntarget. Integrate its result against the actual caller before closing the ledger.\nDo not let parallel workers silently redefine shared objects to suit their proofs.\n\n### Implementation loop\n\nRead [references\u002Flean-working-loop.md](references\u002Flean-working-loop.md) when\nimplementing or repairing an obligation. It develops the search → actual-caller\nexperiment → diagnostic → repair cycle, including representation choices and\nperformance problems, with a compiled library-application example.\n\n1. Select a missing connection or mathematical lemma that changes what the main\n   theorem can prove. State its exact Lean interface and its caller's obligations.\n2. Look for the required results in the pinned library and project. Test unfamiliar\n   declarations locally with `#check`; use existing equivalent representations\n   when they simplify a real bottleneck.\n3. Implement the lemma and an actual use site. Compile the affected module; after\n   integration, compile the downstream target whose status depends on it.\n4. Diagnose the first relevant error. An elaboration\u002FAPI mismatch calls for a\n   local implementation fix. An unavailable assumption calls for its derivation,\n   a different argument, or an explicit remaining mathematical obligation.\n5. When finite reduction becomes expensive, consider a general counting lemma,\n   recurrence or smaller checked certificate. Do not silently enlarge the trust\n   basis merely to make a tactic finish.\n6. At a new load-bearing argument or interface, run checkpoint B. Implement valid\n   fixes, compile them, and request follow-up on the changed obligation.\n7. Update the existing ledger with the declaration, actual caller, verification\n   evidence and remaining premise; then advance to the next connection.\n\nFor example, a theorem of type `Certificate x → Desired x` is useful only after\nthe project constructs `Certificate x` for every original input it needs. Closing\nthat construction and connecting the caller is a distinct result from proving\nthe conditional theorem. Kernel checking does not discharge a parameter merely\nbecause its type has a reassuring name.\n\nDo not use a timer, repeated unchanged builds, or repeated model calls as a proxy\nfor progress. After a concrete failure, try another justified representation or\nproof path; preserve the exact blocker if the requested work cannot yet finish.\nDistinguish an implementation failure from a failed intermediate claim and an\nobstruction to the whole method. Retire an unsuccessful route with its reason;\nrejecting that route does not refute the original theorem. Parallel exploration\nis most useful when the proposed approaches can fail for different reasons.\n\n## 3. Use computation with a proved interpretation\n\nDistinguish finite examples, exhaustive computation over a proved domain,\nchecked certificates, and symbolic\u002Fgeneral proof. A bounded search is evidence\nabout its searched inputs. Exhaustive finite verification can be a proof when\ncoverage, encoding correspondence, and the verification procedure are established\nand the user's constraints permit it. For an end-to-end Lean result, the coverage\nand interpretation must themselves lie in the proved dependency chain under the\ndeclared foundations. A comment claiming that a finite list is exhaustive does\nnot turn checked list entries into a universal Lean theorem.\n\nFor numeric arguments, use exact arithmetic where the claim needs exactness.\nProve the connection between actual objects\u002Fevents and the finite data before\nusing the numeric conclusion. Document conventions that matter: ordered versus\nunordered pairs, multiplicities, indices, zero cases, and rounding.\n\nAgree on the intended trust basis from the specification and project conventions.\nDo not silently introduce axioms or compiler-backed computation to make an\notherwise incomplete kernel-level proof appear complete. Conversely, do not turn\none project's prohibition on enumeration or `native_decide` into a universal\nban. Explain any extra trust assumption and whether it satisfies this task.\n\n## 4. Review the mathematical obligations that remain uncertain\n\nRun A, B and C as described in the core workflow. Read\n[references\u002Fadversarial-review.md](references\u002Fadversarial-review.md) for stage\ntriggers, source packets, adversarial prompts, objection handling, resumption and\nstopping rules. These are mathematical checks, not votes on whether to trust the\nexecutor's summary.\n\nUse `proof-checker` for the existing ARIS proof-audit\u002Fsubmission role. This skill\ndoes not replace its artifact or reviewer contract. When invoked by a parent\nproof audit, return the checked scope and evidence to that run; do not recursively\ninvoke `proof-checker`. A standalone Lean task does not automatically require a\nseparate paper-audit workflow. Where installed, use\n`research-review` \u002F `auto-review-loop` for targeted discussion and repair, with\nthe current host's reviewer routing.\n\nFor direct model consultations, use available authorized tools and their actual\ncontracts. A Grok or Antigravity MCP consultation is not automatically an ARIS\nreviewer overlay. Do not hard-code vendor availability, quotas, retry counts,\nproxy settings, or a particular model into the portable workflow.\n\nAsk concrete questions: translate this declaration; discharge these caller\nhypotheses; find a counterexample to this new lemma; check this reduction's\ncoverage. Preserve raw responses, actual reading\u002Fexecution scope, and the\nexecutor's disposition of each issue. Model agreement is not a proof. Same-family\nopinions remain provisional; cross-family opinions are additional review\nevidence, not mathematical truth certificates.\n\nAn unread artifact, an unanswered question, a failed call or an empty response\nprovides no substantive review verdict. A plan review is not a final-code review;\nreading existing logs is not executing a build. Preserve which claim\u002Fversion was\nactually examined. Retry only within current authorization and provider rules;\ndo not rewrite failures as passes after a later success.\n\nClose objections with a specific derivation, verified witness, correction or\njustified rejection. Preserve unresolved disputes. Apply an invoked ARIS loop's\nown round\u002Fcontinuation contract; do not create an outer polling loop around it.\nFor direct consultations, re-review changed obligations and stop when settled or\nwhen a real availability\u002Fbudget limit is reached. Unresolved mathematics remains\nopen; an unavailable requested review remains pending. Neither a reviewer score\nnor exhaustion of rounds is a proof-completion rule.\n\n### Host execution and review\n\nRead the local host's [reviewer routing](..\u002Fshared-references\u002Freviewer-routing.md)\nfor its current model\u002Feffort policy and tool contracts. Use the A\u002FB\u002FC mathematical\npackets in this skill; they do not require a paper, publication score, or a new\nreview service. For delegated proof work, follow the applicable\n[fan-out convention](..\u002Fshared-references\u002Ffan-out-pattern.md), assign stable lemma\ninterfaces, and integrate each result at its real caller.\n\nUse Codex's available native shell, file, and collaboration tools for Lean work.\nA second Codex agent is useful for decomposition, implementation, or provisional\nreview; record `review_independence: same-family` and\n`acceptance_status: provisional`. It does not fulfill a requested cross-family\ncheckpoint.\n\nFor A\u002FB\u002FC, use the user's selected authorized cross-family reviewer, or an\nalready configured cross-family route: Claude review MCP, Grok MCP, or Gemini\nthrough its review bridge or Antigravity. An installed `research-review` overlay\nmay carry the packet if its review contract fits; the base Codex version remains\nsame-family. Existing overlays cover only their listed skills and do not\nimplicitly override `lean-formalize`. Direct checkpoint calls need no new overlay;\nread [review transport notes](references\u002Fadversarial-review.md#review-transport-and-source-access)\nfor the actual file-access and continuation rules. Do not invent new global\nreviewer flag values. If no cross-family route is available, continue useful\nlocal Lean work and record the affected checkpoint as pending.\n\n## 5. Verify the actual final target\n\nUse [references\u002Fverification-and-handoff.md](references\u002Fverification-and-handoff.md)\nwhen preparing the final build, trust audit, or portable package.\n\nFor completed delivery, establish the applicable evidence below. Keep a proved\ntheorem distinct from any still-pending reproduction or review requirement:\n\n- **The proposition has a proof:** inspect the exported declaration's actual\n  type, not its keyword. `theorem`, `lemma`, or `def` may supply a proof term of\n  the target proposition. Merely defining that proposition, proving `P → P`, or\n  proving an unrelated proposition does not discharge the original target.\n- **Statement alignment:** its expanded type and definitions express the original\n  claim; no internal profile, certificate-validity, coverage, or success premise\n  is left for the user to supply unless the original claim included it.\n- **Actual assembly:** reductions, cases, witnesses, and return to the original\n  conclusion are connected, with all caller hypotheses discharged.\n- **Executed verification:** explicitly build the module that proves the target\n  and print its type and transitive axioms. For a multi-module project, also\n  independently import the actual entry point to check its exported result.\n  Direct compilation plus type\u002Faxiom output suffices for a small self-contained\n  file; do not build a new project just to create another import. Save commands,\n  exit status, and output. A successful build of a different root module does\n  not count.\n- **Trust accounting:** no `sorryAx` or unapproved extra axiom lies on the target's\n  dependency chain. Explain the actual dependencies against the intended basis;\n  the required list need not be exactly any fixed set of axiom names.\n- **Reproducibility:** retain the relevant source and pinned environment, and\n  provide commands that another reader can run.\n- **Requested review:** C concerns the final exported type, definition bodies,\n  axiom output and original claim, not an earlier plan. Record each required\n  reviewer's substantive findings or unavailability separately.\n\nThese checks establish the formal result under its stated foundations. A prose\nproof or HTML edition requires its own correspondence check; Lean does not certify\nevery sentence in a separate exposition. An explicitly required import audit,\nexternal reproduction, or review may remain pending even after the formal theorem\nis checked. Report that missing evidence as such, not as a newly discovered\nmathematical gap.\n\nAn illustrative audit can distinguish `available_interface : Target → Target`\nfrom `original_proof : Target`, even if both compile without extra axioms. The\nfirst leaves the original claim open; the second supplies its proof when `Target`\nmatches the user's statement. A failed reviewer call changes neither type.\n\n## 6. Persist and hand off the evidence actually obtained\n\nReuse the existing specification, proof ledger, and continuation file. A short\nprogress report plus Lean sources, lockfiles, an audit entry point, and logs is\nusually enough. Add new files only for a concrete reader or execution need.\n\nReport distinct facts rather than a single ambiguous PASS:\n\n- Original theorem: checked, partial, changed claim only, or disproved by a\n  verified counterexample; name the declaration and any remaining obligation.\n- Statement alignment and the actual logical trust basis.\n- Reproduction: existing\u002Fincremental build, clean project rebuild, or another\n  person's independent reproduction—whichever actually happened.\n- Reviews: who read what, who ran code, and any requested review still pending.\n\nFor unfinished work, save the smallest remaining mathematical\u002FLean obligation,\nits known dependencies, the last useful error, and the next action. For completed\nwork, record completion so the next session does not revive obsolete gaps.\n\nUse a compact continuation note rather than a second state system:\n\n```text\nOriginal target and source:\nCurrent exported declaration \u002F main module:\nClosed connections and actual build evidence:\nNext unclosed obligation or completion:\nA\u002FB\u002FC reviews: source scope, reviewer, raw record, remaining issue:\nNext action:\n```\n\nResolve current state from actual artifacts, not an old narrative saying either\n“finished” or “not finished.” On resume, check what changed before repeating work.\nDo not merge historical and current reviewer scopes into a larger claimed audit.\n\nIf a research wiki is active, update the affected claim through the existing\n`research-wiki` \u002F `proof-checker` workflow with precise evidence and scope; do not\noverwrite its generated graph or invent a second acceptance format. Literature\nsearch should resolve a mathematical gap, API question, or relevant prior result,\nnot become an unrelated requirement to exhaust the literature.\n\nFor handoff, provide a reader-oriented proof, source, pinned dependencies,\nindependent audit command, and known trust boundary. Use `render-html` when a\nsingle-file reading edition would help. Mathematical source changes require\nappropriate rebuild and statement\u002Faxiom checks. A theorem or hypothesis changed\nin Markdown or HTML is also a claim change: redo alignment and affected requested\nreview. Layout or wording that preserves the claim needs presentation checks.\nKeep change histories in the research record and present the finished proof\ndirectly. Do not restart completed proof work solely to satisfy a workflow ritual.\n",{"name":5,"description":7},"main","18193d71ea3441d7c033dbf195b8a7c75ac96f75","packages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fskills-codex\u002Flean-formalize",0,"https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science","https:\u002F\u002Fgithub.com\u002FCZX-LBJ","一个没事喜欢随便写点代码的小菜鸡\r\n","2026-10-22 09:48:04",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add CZX-LBJ\u002Fdeepseek-science --skill lean-formalize",{"tool":38,"command":39},"bunx","bunx skills add CZX-LBJ\u002Fdeepseek-science --skill lean-formalize",{"tool":41,"command":42},"pnpm","pnpm dlx skills add CZX-LBJ\u002Fdeepseek-science --skill lean-formalize","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fskills-codex\u002Flean-formalize\u002FSKILL.md",true,[46,52,56,60],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"agents\u002Fopenai.yaml","REFERENCE","参考文档",261,"https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fskills-codex\u002Flean-formalize\u002Fagents\u002Fopenai.yaml",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Fadversarial-review.md",14165,"https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fskills-codex\u002Flean-formalize\u002Freferences\u002Fadversarial-review.md",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"references\u002Flean-working-loop.md",10625,"https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fskills-codex\u002Flean-formalize\u002Freferences\u002Flean-working-loop.md",{"path":61,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":62,"sourceHtmlUrl":63},"references\u002Fverification-and-handoff.md",6332,"https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fskills-codex\u002Flean-formalize\u002Freferences\u002Fverification-and-handoff.md",[65,72,79,86,93,100],{"skillNo":66,"name":67,"nameZh":6,"description":68,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M7FEEKSNMRMKW6JEK93PFDFY","bio-free-energy-calculations","Performs alchemical free-energy calculations including relative binding free energy (RBFE \u002F FEP+) and absolute binding free energy (ABFE) via OpenFE, FEP+, GROMACS, AMBER pmemd, and OpenMM with explicit lambda scheduling, soft-core potentials, MBAR\u002FBAR analysis, cycle-closure validation, and protocol-appropriate enhanced sampling. Compares ML alternatives (Boltz-2 affinity, DeepDock). Use when ranking analogs by binding affinity beyond docking accuracy, performing prospective lead optimization, or validating SAR predictions.","https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Fbioskills\u002Fchemoinformatics\u002Ffree-energy-calculations\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M7FEE5CTE5QK55CEMXNNVWY9","bio-conformer-generation","Generates 3D conformer ensembles using RDKit ETKDGv3 with knowledge-enhanced distance geometry, MMFF94\u002FUFF force-field optimization, CREST + GFN2-xTB semi-empirical refinement, and macrocycle-aware torsion preferences. Provides explicit decision rules for single vs ensemble conformer use, RMSD pruning, energy windows, conformer count, and force-field choice. Use when preparing 3D ligands for docking, generating descriptor input for 3D QSAR, or sampling macrocycle\u002Fpeptide conformational ensembles.","https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Fbioskills\u002Fchemoinformatics\u002Fconformer-generation\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M7FEDNGFEMFWD1MSEDR5H7CY","bio-admet-prediction","Predicts ADMET properties using ADMETlab 3.0 (119 platform features, including 77 prediction models with modeled-endpoint uncertainty), ADMET-AI, DeepChem MolNet, and chemprop D-MPNN with explicit handling of OECD QSAR principles, applicability domain assessment, calibration, hERG\u002FCYP\u002FAMES endpoints, and PAINS \u002F Lipinski \u002F Ro5 \u002F Veber \u002F BBB druglikeness filters. Use when filtering compounds for drug-likeness, prioritizing leads by predicted safety, or building an in-house ADMET QSAR model.","https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Fbioskills\u002Fchemoinformatics\u002Fadmet-prediction\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M7FED69GFB54FRF7AV5066DN","writing-systems-papers","Paragraph-level structural blueprint for 10-12 page systems papers targeting OSDI, SOSP, ASPLOS, NSDI, and EuroSys. Provides page allocation, paragraph templates, and writing patterns. Use when user says \"写系统论文\", \"systems paper structure\", \"OSDI paper\", \"SOSP paper\", or wants fine-grained structural guidance for a systems conference submission.","https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fwriting-systems-papers\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":19},"01M7FECVVDWZ0X0VEDTCE4B4DZ","wiki-enrich","Fill in the per-paper TODO sections of research-wiki\u002Fpapers\u002F\u003Cslug>.md pages that literature-ingest skills leave as bare scaffolds. Use when user says 'enrich wiki', 'fill paper TODOs', 'wiki body 補完', '把 paper 摘要寫進 wiki', 'research-wiki 自動填', or after a batch ingest that left papers\u002F as TODO scaffolds.","https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fwiki-enrich\u002FSKILL.md",[],[],{"skillNo":101,"name":102,"nameZh":6,"description":103,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":19},"01M7FECD4BCJ2MTSBDDSE3ZAJ5","web-debug-search","Search GitHub, Stack Exchange, Chinese technical communities, official documentation, and general developer web sources for software errors, compatibility problems, API usage questions, and real-world workarounds. Use for debugging and discovery only; results are not paper-citation evidence.","https:\u002F\u002Fgithub.com\u002FCZX-LBJ\u002Fdeepseek-science\u002Fblob\u002Fmain\u002Fpackages\u002Fskill\u002Fskill-science\u002Fassets\u002Fsources\u002Faris\u002Fskills\u002Fweb-debug-search\u002FSKILL.md",[],[]]